{"id":"c3e00361-f08c-4be7-a863-b397e3a3902d","arxiv_id":"2607.04144","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Semantic block matching via DINOv2 plus selective high-frequency refinement yields higher-fidelity, multi-view-consistent object removal inside 3D Gaussian Splatting than prior one-shot Gaussian or NeRF inpainters.","lead":"This paper presents a 3D Gaussian Splatting pipeline that removes unwanted objects from multi-view scenes by matching semantic blocks across views with DINOv2 and then progressively refining only the blurry subregions. It matters for AR/VR, robotics, and content tools that need clean, consistent scene edits without full re-capture.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"DINOv2 block matching may retrieve semantically similar but geometrically incorrect content; the paper never validates match accuracy or its necessity for the reported gains.","rationale":"The Reader correctly isolates the unvalidated DINOv2 cosine-matching assumption as the weakest link. My concern sharpens the same point: without a direct measurement of match correctness or a control that severs retrieval from mere extra conditioning, the causal contribution of SBM to multi-view coherence remains unproven. The concrete test above would settle the issue with existing public data and no new training. Because the paper already shows consistent metric gains and no internal contradiction, the appropriate stance remains CONDITIONAL (pending the proposed check and code release), not a harsher rejection. No stronger load-bearing flaw (derivation error, circular evaluation, etc.) is present.","tokens_in":13479,"tokens_out":568,"duration_ms":5679,"concrete_test":"On the public SPIn-NeRF and Mip-NeRF 360 scenes, extract the (v*,j*) pairs chosen by Eq. 9 for every target block, project the matched source block into the target camera, and compute (a) mean IoU / Chamfer distance against ground-truth visible geometry and (b) LPIPS between the matched patch and the true background. Re-run the full pipeline after replacing DINOv2 retrieval with random same-size blocks from other views. If match geometric error is high or if random conditioning recovers >70 % of the SBM gain in Table III, the load-bearing assumption fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on SBM (Section IV-B) producing higher multi-view coherence than one-shot baselines. SBM selects source blocks solely by cosine similarity of DINOv2 features of s\times s patches (occlusion threshold \tau=0.1) and injects the best-match token via cross-attention (Eqs. 8–10). Nothing in the paper measures whether those matches are geometrically or texturally correct: no match-accuracy metric, no qualitative match visualizations, no failure cases under viewpoint change or repetitive texture. Ablation Table III only removes the entire SBM module; it does not isolate the retrieval step from the mere presence of extra conditioning. Consequently it remains possible that the observed PSNR/SSIM/LPIPS gains (Tables I–III) arise from additional diffusion conditioning or from RPR alone rather than from correct cross-view geometric alignment. The hand-tuned thresholds (\tau=0.1, 15 % high-frequency) further leave open the possibility that the method is brittle outside the reported scenes.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a 3D object-removal pipeline built on 3D Gaussian Splatting that first obtains an initial completion via SDS from a latent diffusion model, then applies Semantic-guided Block Matching (SBM) and Region-Wise Progressive Refinement (RPR). SBM partitions the masked region into s\times s blocks, extracts DINOv2 features, retrieves the highest-cosine-similarity block from other views (Eqs. 8–10), and injects the matched token via cross-attention into the UNet. RPR flags low-quality blocks by a Sobel high-frequency magnitude threshold (15 % of surroundings) and re-diffuses only those blocks with both semantic and frequency conditioning (Eqs. 11–14). The refined multi-view images supervise the final Gaussian optimization under a combined L_GS + λ_SDS L_SDS loss. Quantitative results on SPIn-NeRF, self-captured, and Mip-NeRF 360 scenes (Table I) and ablations (Table III) show gains in masked-region PSNR/SSIM/LPIPS over SPIn-NeRF, GaussianEdit, GaussianGroup, MVInpainter and InFusion, with competitive runtime/memory (Table II).","tokens_in":13788,"tokens_out":1257,"duration_ms":9373,"significance":"If the reported gains hold under broader validation, the work supplies a practical, relatively efficient route to multi-view-consistent object removal inside the now-standard 3DGS representation—directly useful for AR/VR editing, robotics scene cleaning, and digital content pipelines. The explicit combination of DINOv2 block retrieval with selective high-frequency re-diffusion is a concrete engineering contribution beyond pure one-shot SDS or 2D inpainting baselines. Strengths include clear module ablations (Table III), multi-dataset quantitative tables restricted to masked pixels, and a favorable efficiency profile (7.9 h / 10 GB). The absence of match-accuracy diagnostics and statistical error bars limits the strength of the causal claim for SBM, but does not erase the empirical utility of the overall system.","major_comments":[{"comment":"Section IV-B (Eqs. 8–10) and the central claim of multi-view geometric coherence rest on the assumption that cosine similarity of DINOv2 features of s\times s blocks (τ = 0.1) selects geometrically and texturally correct source content. The manuscript never reports match accuracy, qualitative match visualizations, or failure cases under viewpoint change or repetitive texture. Ablation Table III only removes the entire SBM module; it does not isolate retrieval correctness from the mere addition of extra conditioning tokens. Without such evidence it remains possible that the PSNR/SSIM/LPIPS gains in Tables I–III arise from richer diffusion conditioning rather than correct cross-view alignment. A modest diagnostic (e.g., fraction of matches whose 3D re-projection lands inside the corresponding unmasked region, or a controlled ablation that replaces the best-match token by a random same-view","section":"Section IV-B, Eqs. 8–10; Table III"},{"comment":"The free parameters that control both modules—occlusion threshold τ = 0.1, high-frequency quality threshold δ = 15 % of surroundings, block size s = 15, and SDS weight/iteration count—are stated as empirical choices (Sections IV-B.1, IV-C.1, V-A.1) with no sensitivity analysis. Because the reported superiority over one-shot baselines depends on these thresholds, a short sensitivity study (or at least results for two neighboring values of τ and δ) is required to show that the gains are not brittle to the particular operating point chosen on the evaluation scenes.","section":"Sections IV-B.1, IV-C.1, V-A.1"}],"minor_comments":[{"comment":"Figure 1 caption and the surrounding text refer to a “High-frequency feature extractor” and “frequency-aware UNet,” yet the precise architecture of E_freq (Eq. 12) and how f_k is fused into the UNet (cross-attention vs. MLP) are never specified beyond a one-sentence description.","section":"Section IV-C.2, Fig. 1"},{"comment":"Table I reports point estimates without standard deviations or statistical significance tests across scenes; adding per-scene variance or a paired test would strengthen the quantitative claims.","section":"Table I"},{"comment":"Notation for the semantic encoder switches between F_DINOv2 (Eq. 8) and F_sem (text after Eq. 8); a single consistent symbol would improve readability.","section":"Section IV-B"},{"comment":"The abstract and introduction claim “outperforms existing Gaussian-based methods,” yet the related-work discussion of GScream and GaussianEditor is brief; a short paragraph clarifying the precise technical differences would help readers place the contribution.","section":"Section II-B"},{"comment":"Minor typographical issues: “Region-Wise Progressive Refine- ment” line break, “Fe-map&mask” in Fig. 1, and inconsistent capitalization of “High-frequency.”","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The technical novelty is incremental but solid engineering; the main risk is that the geometric-alignment story for SBM is currently under-supported. If the authors can supply even a lightweight match-accuracy diagnostic and a short sensitivity plot, the paper becomes a clear accept for a solid systems venue. Without those additions the causal claim for SBM remains the weakest link."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean incremental systems paper for 3DGS object removal. The two modules they actually introduce—SBM (DINOv2 block retrieval + token injection into the UNet) and RPR (Sobel high-frequency thresholding + selective re-diffusion)—are not in GaussianEditor, InFusion, MVInpainter or GaussianGroup. Tables I–III and the ablations show consistent PSNR/SSIM/LPIPS lifts on masked regions across SPIn-NeRF, Mip-NeRF 360 and their own captures, and the runtime/memory numbers are competitive (7.9 h / 10 GB). That is real, usable engineering for AR/VR and robotics pipelines that already live in 3DGS.\n\nWhat they do well: the pipeline is readable, the losses are standard (L1 + D-SSIM + SDS), the ablations isolate each module, and they report both forward-facing and unbounded scenes. No circular math; metrics are measured against external baselines.\n\nSoft spots, in proportion. The stress-test is right that they never measure whether the DINOv2 cosine matches are geometrically correct—no match accuracy, no visualizations, no failure cases under viewpoint change or repetitive texture. Ablating the whole SBM module does not isolate retrieval quality from “extra conditioning.” The free parameters (τ=0.1, 15 % high-freq, s=15) are purely empirical with no sensitivity. Private data and missing code also limit reproducibility. These are real gaps, but they do not collapse the central claim that the combined pipeline beats the cited one-shot baselines on the reported metrics.\n\nWho it is for: people already shipping 3DGS editing tools. Not a new scientific regime, but a legitimate advance that a serious referee should see. I would accept for peer review, push for match-quality diagnostics and code, and expect the paper to survive with those additions.","headline":"Solid engineering paper: two concrete modules (DINOv2 block matching + frequency-selective re-diffusion) give measurable gains over recent Gaussian inpainters; the match-correctness claim is under-validated but not fatal.","tokens_in":14378,"tokens_out":489,"would_cite":true,"duration_ms":4475,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Semantic block matching plus selective high-frequency re-inpainting removes objects from 3D Gaussian scenes more coherently than one-shot methods.","keywords":["3D Gaussian Splatting","object removal","semantic block matching","DINOv2","region-wise progressive refinement","Score Distillation Sampling","multi-view inpainting","high-frequency guidance"],"falsifier":"On a held-out multi-view scene with large occlusions, disable either the DINOv2 matching or the high-frequency selective re-diffusion and measure whether the masked-region PSNR/SSIM/LPIPS gap versus the full pipeline disappears or reverses relative to the one-shot baselines reported in the paper.","tokens_in":14396,"feed_emoji":"🧹","tokens_out":682,"duration_ms":5239,"temperature":0.7,"pith_summary":"This paper claims that 3D object removal inside Gaussian Splatting becomes both sharper and more multi-view consistent when two steps replace the usual one-shot fill of an entire mask. First, DINOv2 embeddings of occluded blocks are matched by cosine similarity to the best-matching visible blocks from other camera views; those tokens are then injected via cross-attention into a diffusion model so the missing content is completed with semantically aligned material. Second, a simple high-frequency (Sobel) score flags only the sub-regions that still look blurry or inconsistent, and those blocks alone are re-diffused with an extra frequency prior. The authors show that the combination raises PSNR, SSIM and LPIPS inside the masked regions on forward-facing and unbounded scenes, while keeping training and memory costs modest relative to competing Gaussian and NeRF baselines. A reader who cares about AR/VR editing or robotics scene clean-up would therefore have a practical route to remove large objects without leaving ghosting or texture seams across views.","feed_headline":"Semantic blocks plus selective re-fill clean 3D Gaussians","feed_subtitle":"Matching DINOv2 tokens across views and re-diffusing only blurry patches beats one-shot object removal.","key_machinery":"Semantic-guided Block Matching (SBM) plus Region-Wise Progressive Refinement (RPR): occluded s\times s blocks are retrieved by DINOv2 cosine similarity and used as semantic tokens; low-frequency residual blocks are then selectively re-inpainted under a high-frequency prior.","core_discovery":"The authors establish that Semantic-guided Block Matching (DINOv2 token retrieval followed by cross-attention injection) together with Region-Wise Progressive Refinement (high-frequency thresholding of sub-masks and selective re-diffusion) produces higher-fidelity, multi-view-coherent object removal inside 3D Gaussian Splatting than existing one-shot Gaussian-based or NeRF-based pipelines.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["DINOv2 tokens match views for progressive Gaussian object removal","Semantic block matching plus selective re-diffusion cleans 3D Gaussians","Region-wise refine on high-freq submasks beats one-shot Gaussian erase","Cross-view DINOv2 guidance enables coherent progressive object cutout","SBM token retrieval and RPR re-fill raise 3D Gaussian removal fidelity"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes that cosine similarity of DINOv2 block features reliably finds the geometrically correct source patch and that a simple Sobel magnitude threshold correctly marks only the blocks that still need re-inpainting.","fun_headline_variants_meta":{"raw":{"variants":["DINOv2 tokens match views for progressive Gaussian object removal","Semantic block matching plus selective re-diffusion cleans 3D Gaussians","Region-wise refine on high-freq submasks beats one-shot Gaussian erase","Cross-view DINOv2 guidance enables coherent progressive object cutout","SBM token retrieval and RPR re-fill raise 3D Gaussian removal fidelity"]},"model":"grok-4.5","effort":"low","cost_usd":0.005914,"raw_usage":{"total_tokens":1517,"prompt_tokens":749,"num_sources_used":0,"completion_tokens":100,"cost_in_usd_ticks":59140000,"prompt_tokens_details":{"text_tokens":749,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":668,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":749,"tokens_out":100,"duration_ms":5626,"temperature":1.0,"reasoning_tokens":668,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T21:22:15.060341+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out multi-view scene with large occlusions, disable either the DINOv2 matching or the high-frequency selective re-diffusion and measure whether the masked-region PSNR/SSIM/LPIPS gap versus the full pipeline disappears or reverses relative to the one-shot baselines reported in the paper.","supporting_citations":[],"review_version":1}